* Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > On Sat, Dec 14, 2013 at 03:19:02PM +0100, Peter Zijlstra wrote: > > On Fri, Dec 13, 2013 at 10:11:05AM +0800, Alex Shi wrote: > > > BTW, > > > A bewitching idea is till attracting me. > > > https://lkml.org/lkml/2012/5/23/148 > > > Even it was sentenced to death by HPA. > > > https://lkml.org/lkml/2012/5/24/143 > > > > > > That is that just flush one of thread TLB is enough for SMT/HT, seems > > > TLB is still shared in core on Intel CPU. This benefit is unconditional, > > > and if my memory right, Kbuild testing can improve about 1~2% in average > > > level. > > > > > > So could you like to accept some ugly quirks to do this lazy TLB flush > > > on known working CPU? > > > Forgive me if it's stupid. > > > > I think there's a further problem with that patch -- aside of it being > > right from a hardware point of view. > > > > We currently rely on the tlb flush IPI to synchronize with lockless page > > table walkers like gup_fast(). > > > > By not sending an IPI to all CPUs you can get into trouble and crash the > > kernel. > > > > We absolutely must keep sending the IPI to all relevant CPUs, we can > > choose not to actually do the flush on some CPUs, but we must keep > > sending the IPI. > > The alternative is switching x86 over to use HAVE_RCU_TABLE_FREE. So if the kbuild speedup of 1-2% is true and reproducable then that might be worth doing. Building the kernel is obviously a prime workload - and given that the kernel is active only about 10% of the time for a typical kernel build, a 1-2% speedup means a 10-20% speedup in kernel performance (which sounds a bit too good at first glance). Thanks, Ingo -- To unsubscribe, send a message with 'unsubscribe linux-mm' in the body to majordomo@xxxxxxxxx. For more info on Linux MM, see: http://www.linux-mm.org/ . Don't email: <a href=mailto:"dont@xxxxxxxxx";> email@xxxxxxxxx </a>